Improve ability to load pretrained models

Currently, we can do this:

```
from nn_lib.models import get_pretrained_model

model = get_pretrained_model("resnet18")
```

This is a wrapper around the `torchvision.models` module. We should look into other model sources. The huggingface `transformers` or `timm` packages should be added. And if `timm` subsumes `torchvision.models` then we should perhaps remove the torchvision interface.